🚧 Development Journey

🧩 Environment Hell

From conflicting Python versions to VS Code interpreter chaos, environment setup consumed our first full day. We eventually consolidated to a single global Python 3.12 environment and reinstalled everything cleanly.

---

⚙️ GPU Integration

We made FinBERT GPU-accelerated by:

• Installing CUDA-enabled PyTorch
• Verifying GPU usage with

  nvidia-smi

• Explicitly binding the model and tensors to CUDA

---

⚡ The FinBERT Challenge: Beating the “Too Slow” Problem

At first, we assumed AI-driven sentiment analysis would be too computationally expensive for real-time use.
FinBERT, being a large Transformer model, is known for slow inference times on CPU — and we were analyzing up to 100 headlines per ticker over a 30-day window.

Our first prototype on CPU confirmed this: each batch took 20–30 seconds to process, making it unusable.

We overcame this bottleneck through three key optimizations:

1. 🚀 GPU Acceleration (PyTorch + CUDA):
   Offloaded FinBERT to an RTX 4060 GPU, reducing inference time to under 2 seconds per ticker — a 15–20× speedup.

2. 🧠 Sentiment Caching:
   Implemented in-memory caching for repeated headlines to save processing cycles.

3. 📅 Daily Sentiment Aggregation:
   Aggregated daily results instead of headline-by-headline analysis, improving clarity and speed.

These steps turned FinBERT from a performance bottleneck into a real-time sentiment engine.

---

🧑‍💻 Collaboration Chaos: Git for the First Time

Half our team used Git for the first time — during a hackathon.

We faced:

• Merge conflicts on shared files like

  index.html

• Lost commits from incorrect pulls
• Confusion over branching and pushing

We recovered by:

1. Assigning ownership per module (backend, frontend, UI).
2. Learning Git basics together live.
3. Committing small and frequently.
4. Restoring working versions via GitHub history.

By the end, the same team that struggled with Git merges was confidently collaborating in true parallel development.

---

🎯 Frontend Integration: The Chart–Sentiment Bridge Problem

After separate backend and frontend builds, integration became the hardest phase.

🧩 The Early Struggle — Highcharts

We first tried Highcharts, but it wasn’t robust enough:

• Performance drops on large datasets
• Weak overlay support
• Inconsistent OHLC rendering

We switched to TradingView Lightweight Charts, which offered:

• Smooth performance
• Professional-grade rendering
• Perfect control for financial overlays

⚙️ Integration Issues

• Timestamp mismatches
• Async render delays
• Tooltip desync between sentiment and candles

✅ Our Fix

• Unified timestamps in UTC
• Merged data into one payload from Flask
• Deferred rendering until all data fetched
• Synced sentiment highlights with chart hover

This finally gave Sentichart a seamless, responsive experience.

---

💸 The API Limitation Wall

We were restricted to free-tier APIs and libraries, including:

• NewsAPI (100 requests/day, 30-day history)
• yFinance (rate-limited)
• Hugging Face (local-only FinBERT inference)

We adapted by caching data, batching requests, and preloading sentiment for key tickers.

---

🔍 The Search Bar Experiment: UX Meets API Limits

To enhance usability, we built a ticker search bar.
The plan: query live stock suggestions via API and auto-complete.

But due to API limits and broken keys, we couldn’t fetch symbols dynamically.
Instead, we hardcoded popular tickers (AAPL, TSLA, MSFT, NVDA, AMZN, GOOG) for demo use.
The modular system still supports swapping to live API search later — no structural rewrite needed.

---

🐞 Backend Bug Hunt: The Case of the Missing Indicators

Multi-line indicators like Bollinger Bands broke because

pandas_ta

created multiple columns.
We fixed this by detecting all new columns dynamically before returning data to the frontend.

---

⚖️ The Indicator Overload: When More Became Less

Some indicators (like Ichimoku Cloud and SuperTrend) proved unstable and unresponsive with our real-time structure.
We removed them to preserve chart fluidity and ensure consistent redraws.

The result: a stable, high-performance chart focused on reliability and clarity.

---

🔥 The Firewall Fiasco

Our API failed over LAN due to Windows Firewall restrictions.
We resolved it by adding an inbound rule for

python.exe

.

---

🎨 Frontend Evolution: The Quest for Clarity

Our frontend evolved through multiple failed experiments:

• Attempt 1: Text Markers — unreadable clutter.
• Attempt 2: Histogram Bars — clean but detached from price action.
• Attempt 3: Elegant Dots — visually nice but unclear in interaction.
• ✅ Final Solution: A right-hand Sentiment Timeline Panel, combining clarity, detail, and performance.